What is opencode-dataset-plugin?

opencode-dataset is a plugin for OpenCode, an AI coding agent, that captures your agent sessions and turns them into a local training dataset. Every tool call, file edit, message, and permission request that happens during a session gets written to a plain text file in a line by line JSON format called JSONL, which is easy to read, diff, and load into common machine learning tools like the Hugging Face Datasets library. Before anything is saved, the plugin redacts common secrets such as API keys, bearer tokens, GitHub tokens, AWS access keys, JWTs, private keys, and fields named password, token, secret, or authorization. This means the captured dataset should not contain your credentials even if they appeared in a session. Once records are captured, you can review them using a terminal interface built with OpenTUI. It shows pending examples first so you can accept or reject them with single key presses, and switches to showing everything once nothing is left pending. There is also a one shot review mode meant for automated checks in continuous integration pipelines. Accepted records can then be exported into several formats that are commonly used to train or fine tune language models: a simple OpenAI style chat format, the Alpaca instruction format, ShareGPT conversation format, and a preference based DPO format with chosen and rejected examples. The plugin also adds a tool inside OpenCode itself called save_training_example, so you can ask the agent to save a completed task as a labeled training example directly, and that record is marked accepted right away. Installation is a single command through OpenCode's plugin system, followed by a small JSON configuration entry pointing to where the dataset file should live. The project is written in TypeScript and uses Bun for local development, testing, and building.